Cardiovascular diseases prediction by integrated risk factors assessment by means of machine learning

Aim. To develop a model by machine learning to predict the risk of cardiovascular diseases (CVD) and validate the model using Russian medical data. Materials and methods. The data set was obtained from the Framingham study, consisting of 4,363 patients without CVD, 852 (19.5%) of which died of myocardial infarction and stroke within 10 years of observation. Incoming model features: gender, age, systolic blood pressure, cholesterol, smoking, body mass index, heart rate. The original data set was divided into 2 parts: the training data set (80% of the records) and the validate data set (the remaining 20%). Additionally, the model was evaluated by an external data set included 411 depersonalized patient data from the Russian citizens. Results. The WML.CVD.Score model was created by the serial neural network with one input, two hidden and one output layer. Accuracy results on a training dataset: Accuracy 81.15%, AUC 0.80. The same indicators on the validate data set were: Accuracy 81.1%, AUC 0.76. Test results for the test data set: Accuracy 79.07, AUC 0.86. On the Russian test data, the AUC for the SCORE scale was 0.81 versus 0.86 for the developed model, which showed the validity of the use of machine learning in order to increase the predictive model. Conclusion. The developed model has demonstrated high accuracy to CVD predicting in both internal and external validation. The model can be used in medical practice for patients in Russia.

cardiology, cardiovascular diseases, risk factors, risk modeling, machine learning